This article summary is based on the research paper: 'Precipitaion Nowcasting using Deep Neural Network' and Techxplore post. All credits for this research goes to the authors of this paper. 👏 👏 👏 👏 Please don't forget to join our ML Subreddit Need help in creating ML Research content for your lab/startup? Talk to us at [email protected]
Deep learning models are incredibly successful in analyzing massive quantities of data and accurately forecasting future occurrences.
Meteorologists can now fairly accurately forecast broad weather patterns for the next two to three days. However, climate change has increased unexpected extreme weather events such as thunderstorms, hailstorms, and hurricanes. Predicting these unexpected weather phenomena accurately a few hours ahead of time might help people prepare for them, perhaps reducing their effects and negative consequences.
Three deep neural networks have recently been constructed by researchers at IRT AESE Saint Exupéry and Météo-France to anticipate oncoming precipitation. These networks, first described in a study pre-published on arXiv, might help meteorologists, governments, sports event organizers, and other organizations forecast the advent of storms, hurricanes, and other extreme weather phenomena one to six hours ahead time.
The researchers noted in their report, “We suggest using three prominent deep learning models, the three being U-net, ConvLSTM, and SVG-LP, trained on two-dimensional precipitation maps for precipitation nowcasting.” “We also suggested a patch extraction approach for obtaining high-resolution precipitation maps.”
Most long-term weather forecasts now rely on numerical models that use photographs of the sky, radar data, and other accessible atmospheric data to mimic atmospheric physics processes. While these approaches can accurately anticipate precipitation, they frequently need substantial computations and take a long time to perform. As a result, these methodologies may not always perform equally well in precipitation nowcasting or impending forecasting precipitation.
Here are two instances of precipitation forecasts. The network’s findings are made up of six model outputs. Ground truth is on the top row; the U-net model is on the second row; the ConvLSTM model is on the third row, and the SVG-LP model is on the bottom row.
The researchers’ goal at IRT AESE Saint Exupéry and Météo-latest France’s research was to construct deep neural networks that could better deal with precipitation nowcasting than numerical weather forecasting models. They offered three distinct models in their study: a U-net, a ConvLSTM, and an SVG-LP network.
The three networks have been trained on a dataset of 20,352 high-resolution pictures acquired by Météo-France between 2017 and 2018 using radar echo technology. These photos covered a region in France of around 1000 × 1000 km2.
Because feeding the high-resolution precipitation maps straight to deep neural networks would overburden a computer’s GPU, the researchers devised a patch extraction approach that divides them into 256 x 256 patches. Instead of forecasting precipitation for the entire map, the networks may learn to predict rainfall for individual patches. Finally, they devised a loss function technique that enhances the quality of pictures processed by neural networks, making them less fuzzy.
The researchers put the three models they created through a battery of tests, evaluating the quality of the reconstructions they generated and the accuracy of their forecasts. While all three models accurately predicted the evolution of precipitation fields, the U-Net model, which is developed on a convolutional neural network (CNN) architecture, outperformed the other two.
The researchers claimed in their report that “the CNN-based technique beats the RNN-based models.” “It can create high-value precipitation and anticipate future rainfall contours more precisely. ConvLSTM also beats SVG-LP, although it blurs later frames.”
The U-Net architecture created by this group of academics might be utilized in the future to develop more effective systems for forecasting impending precipitation and rainstorms. Furthermore, this study may motivate other groups to create similar models to anticipate extreme weather occurrences.